We’ve seen that Artificial intelligence (AI) is useful for everyday tasks: managing your agenda, writing emails or summarizing documents has become much easier.

But AI has the potential to be a powerful innovator – and potentially save lives. In a recent proof of concept, we developed a smart bot with Azure AI Studio and GPT-4 that cut waiting times in half. In this article, we will cover how we approached the problem. 

More emergency calls, less staff

Health care institutions struggle to keep health care affordable and attractive. Increasing demand for care, the rising costs and a shortage of staff are noticeable in the emergency department – they have to respond quickly and accurately to emergency calls, and assess how serious the situation is. Every second counts. 

A typical process for emergency calls looks like this:

But what if there are more and more calls, and the waiting times get longer?
What if the staff gets overwhelmed and makes mistakes?
And what if, and this is the core of the problem, there is little specialized staff available to process and analyze the calls?

That’s what emergency care struggles with. And that’s where artificial intelligence (AI) can make a difference. 

Building the process with Azure AI Studio and GPT-4

We saw an opportunity to tackle this issue, that many regions are struggling with.

With the help of Azure AI Studio and GPT-4, we created a bot that supports the staff by answering questions. Our approach was based on prompt engineering to train GPT-4, according to existing guidelines.

The goal was to enable the bot to smart decisions about the severity of the emergency calls. This required to feed the model with a series of carefully crafted prompts and answers. That effectively taught it to classify emergencies accurately. In addition, we integrated Azure AI Speech Service, so that the AI agent could not only speak, but also understand Dutch (in different dialects). 

The reality check

When a person in distress (or someone nearby) calls, the bot uses specific questions to evaluate the urgency of the situation.

For example, when a person is experiencing shortness of breath, the digital assistant asks control questions.

In addition, the bot also detects whether someone can communicate well, is panting, etc.

In this way, the urgency is determined. An ambulance is immediately dispatched and advice is given on further assistance. Or, if an ambulance is not necessary, the bot asks further questions and advises within what time frame a doctor should be consulted or whether self-help is sufficient. 

The AI solution also automatically generates a report at the end of each conversation, stating the situation, the conclusion and the decision. This helps in creating records for auditing purposes. 

Smart bot, big impact: 50% reduced waiting times in the emergency department 

The outcomes from the testing phase with the AI agent are promising and show clear benefits that not only enhance operational capabilities, but also have a direct impact on the quality of care.

Here is an overview of some of the measurable successes and advantages that we have observed: 

• 50% shorter waiting times: Our solution reduced waiting times by more than 50%, a significant improvement that can save lives in critical situations. 
• Consistent classification: Although the bot’s classification accuracy is not yet flawless, it consistently follows the guidelines provided, ensuring reliability in the decision-making process. 
• Scalability: The AI agent can handle a larger volume of calls without compromising performance. 
• Improved interpretation of clinical data: AI’s fast inference abilities improve the interpretation of clinical data, helping to prioritize urgent cases. 
• Reduced workload: The most important result is that staff experience less workload and stress, allowing them to focus on critical tasks where human expertise is indispensable. And they enjoy their work more. 

Key factors to make AI bots in health care a success

As we learned, the bot offers unprecedented opportunities for improving patient triage and optimizing care pathways. However, implementing AI in healthcare is not without challenges. The success of AI depends on a number of key factors: 

• Definition of triage guidelines: To work effectively, the symptoms and conditions that will guide patients to the right outcome must be clearly defined. 
• User training: End users, such as emergency physicians, need to be trained to enter data correctly into the AI tool so that the tool can make accurate triage decisions. 
• Ethical challenges: AI is only as good as its creator, who may have biases. Although Azure AI Studio helps filter these, some biases are necessary, for example genetic abnormalities, which the AI assistant does not yet take into account. 
• Input quality: It is essential that guidelines and input are done correctly, and that the patient’s background is taken into account. In healthcare, it is crucial to minimize errors. That is why we advise using the bot as an assistant, to support the conversation, rather than as a substitute for people. 

What’s next? 

The results of the AI solution are promising, but much more is possible. A next step is to input the conversation reports, so that AI can analyze patterns in patient data, identify people at risk of certain conditions, and recommend preventive measures. This ultimately reduces the pressure on emergency services. 

The virtual assistant can be used in various ways and can take over any form of interaction based on guidelines. For example, the bot can help general practitioners with giving advice and making diagnoses. During virtual consultations, AI can improve diagnostic accuracy and streamline documentation.